Evaluation of Feature Detectors for KLT based Feature Tracking using the Odroid U3

Feature tracking is an integral part of most vision-based state estimation frameworks. However, tracking features at a sufficient frame rate is a challenging task for mobile robots such as Micro Aerial Vehicles (MAVs) due to their fast dynamics and limited on-board computing resources. Recent developments in smartphone processors have led to embedded computing platforms that are ideal on-board computers for MAV state estimation. This paper analyses the performance of a Kanade-Lucas-Thomasi (KLT) based feature tracker on a state-of-theart embedded computing platform suitable for on-board MAV state estimation. It compares the performance of different implementations of the feature tracker using four different lowcomplexity feature detectors. The experimental results presented herein may serve as guidelines for the selection of a feature detector, image resolution, framerate and feature quantity when developing on-board feature tracking systems based on ARM Cortex-A9 embedded computers.